Colour-based dermoscopy classification of cutaneous lesions: An alternative approach

• Type: Article in International Scientific Journal
• Date: 2013

Authors

Silva, CSP (Author) Other; marcal, ars (Author) FCUP

Journal

Title: Computer Methods in Biomechanics and Biomedical Engineering: Imaging and Visualization

Vol. 1

Pages: 211-224

ISSN: 2168-1163

Publisher: Taylor & Francis

Indexing

Scopus - 10 Citations

Other information

Authenticus ID: P-00K-R9R

DOI: 10.1080/21681163.2013.803683

Abstract (EN): Dermoscopy (dermatoscopy or epiluminescence microscopy) is a non-invasive diagnostic technique for the in vivo observation of pigmented skin lesions used in dermatology. There is currently a great interest in the prospects of automatic image analysis methods for dermoscopy, both to provide quantitative information about a lesion, which can be of relevance for the dermatologist, and as a stand-alone early warning tool. The standard approach in automatic dermoscopic image analysis has usually three stages: (i) segmentation, (ii) feature extraction and selection and (iii) lesion classification. This study evaluates the potential of an alternative approach based on the Menzies method - presence of one or more of six colour classes, indicating that the lesion should be considered a potential melanoma. This method does not require stages (i) and (ii) - lesion segmentation and feature extraction. The identification of colour classes in dermoscopic images is a subjective task, which poses great challenges for an automatic implementation. The purpose of this work is to evaluate the potential discrimination between the six Menzies colour classes in dermoscopic red, blue and green (RGB) images. The Jeffries-Matusita and transformed divergence separability distances were used to evaluate the colour class separability for an experimental evaluation with 28 dermoscopic images. Considering the skin as an additional class, an image intensity calibration was applied to the data-set, which improved the rate of separable colour class pairs. A nonlinear cluster transformation allowed almost the total separation of each colour class in the feature space. Several neural networks in competition were used as classifiers, which lead to loss of arbitrariness and perfect knowledge of each cluster surface. The discrimination between the various Menzies colour classes in dermoscopic RGB images achieved 93% of sensibility, 62% of specificity and 74% of accuracy (averaged measures). These results indicate that it might be possible to evaluate a lesion based on the presence of Menzies colours in dermoscopic images, mimicking the human diagnosis. © 2013 Copyright Taylor and Francis Group, LLC.

Language: English

Type (Professor's evaluation): Scientific